Abstract
Music source separation, as a fundamental task in intelligent audio processing, plays a critical role in enhancing the performance of music generation, editing, and understanding systems. However, existing separation models often suffer from structural limitations such as reliance on a single modeling path, entangled time-frequency representations, and difficulty in adapting to heterogeneous sound sources. Furthermore, they struggle to maintain an effective balance between long-range dependency modeling and inference efficiency. To address these challenges, this paper proposes a novel dual-path state space modeling architecture, MSNet. By introducing decoupled modeling mechanisms for temporal and frequency pathways, and incorporating Mamba-based state space units for multidimensional structural parsing of audio signals, MSNet enhances selective control and structural representation in time-frequency modeling. Experimental results demonstrate that MSNet achieves state-of-the-art performance on the MUSDB18 dataset across multiple evaluation metrics. In particular, it shows superior robustness and stability when dealing with dynamically complex sources such as vocals and drums. Additionally, the model achieves a real-time factor (RTF) below 0.1 while maintaining superior separation quality, making it suitable for deployment in practical applications. This study not only demonstrates the feasibility of state space models for complex audio modeling but also introduces a new architectural paradigm for music source separation that balances accuracy and efficiency. The implementation is publicly available at: https://github.com/NMLAB8/Mamba-S-Net.